Global Contrast Enhancement Detection via DeepMulti-Path Network

Cong Zhang, Dawei Du∗, Lipeng Ke, Honggang QiSchool of Computer and Control Engineering

University of Chinese Academy of Sciences, Beijing, ChinaEmail: zhangcong171@mails.ucas.ac.cn

Siwei LyuComputer Science Department

University at Albany, SUNY, Albany, USAEmail: slyu@albany.edu

Abstract—Identifying global contrast enhancement in an imageis an important task in forensics estimation. Several previousmethods analyze the “peak-gap” fingerprints in graylevel his-tograms. However, images in real scenarios are often stored inthe JPEG format with middle/low compression quality, resultingin less obvious “peak-gap” effect and then unsatisfactory perfor-mance. In this paper, we propose a novel deep Multi-Path Net-work (MPNet) based approach to learn discriminative featuresfrom graylevel histograms. Specifically, given the histograms,their high-level peaks and gaps information can be exploitedeffectively after several shared convolutional layers in the net-work, even in middle/low quality compressed images. Moreover,the proposed multi-path module is able to focus on dealingwith specific forensics operations for more robustness on imagecompression. The experiments on three challenging datasets (i.e.,Dresden, RAISE and UCID) demonstrate the effectiveness of theproposed method compared to existing methods.

I. INTRODUCTION

With the rapid development of digital imaging devices,it is important to verify the authenticity of digital images,as they can be easily manipulated using graphics editingsoftware (e.g., Adobe Photoshop). Detecting global contrastenhancement can provide important cues on the authenticityof a digital image. Contrast enhancement is a nonlinearfunction of pixel values that changes the overall distributionof the pixel intensities in a image, such as gamma correction,sigmoid stretching and histogram equalization. Global contrastenhancement operations may not be the direct result of tam-pering, but many image forgeries involve such operations tohide the traces of other tampering operations such as regionsplicing or cloning.

When a contrast enhancement operation is applied to animage, its pixel values and histogram bins undergo a nonlinearmapping that leaves a distinct “peak-gap” effect on the pixelhistograms. These peaks and gaps, an example can be seen inFig. 1(b) and (c), can be used as fingerprints to identify thecontrast enhancement operation [1]–[4]. These methods workwell for uncompressed images, but their performances are notsatisfactory when the digital images are in the JPEG format [5]with middle/low compression quality. Particularly, low qualityJPEG compression corresponds to a smoothing of pixel values,which weakens peak and gap bins in graylevel histogramsand make them smooth again (see Fig. 1(d) and (f)), resulting

Dawei Du is the corresponding author.

Fig. 1. Examples of global contrast enhancement operations.

in failures for existing global contrast enhancement detectionmethods. Furthermore, as seen in Fig. 1(c) and (e), the multiplesource of forensics operations (e.g., gamma correction andhistogram equalization) usually has different kinds of peak-gap distributions. It is also difficult to detect them effectivelywith fixed rule based thresholds [3].

To solve these issues, in this paper, we propose a deepMulti-Path Network (MPNet) based method to detect globalcontrast enhancement operations. In contrast to existing meth-ods, our network can exploit more discriminative features inconsecutive bins of graylevel histograms to represent peaksand gaps, even when the original peak and gap features havebeen diminished by JPEG compression. Moreover, to detectdifferent forensics operation effectively, we develop a multi-path module in the network, in which each path is to learnrepresentations for a specific contrast enhancement operation.Ahead of these paths, we use several shared convolutional

Fig. 2. Architecture of (a) VGG network and (b) the proposed multi-path network. The red and green path are used to learn features from gamma correctionand histogram equalization operation, respectively.

layers to capture common low-level features of histograms.Notably, the back-propagation training algorithm does notdistinguish these paths from each other, and thus cannotguarantee each path to learn specific representation for thecorresponding forensics operation. Therefore we employ amulti-stage training method for the proposed network. First,we train the shared layers using VGG network; then we trainthe paths individually based on samples of correspondingoperation types; finally we combine the multiple paths andoutput the label whether the image is altered. The contributionsof this paper are summarized below:

• We propose a novel deep Multi-Path Network (MPNet)for global contrast enhancement detection, where eachpath is used to learn specific forensics representation.

• We develop a multi-stage training method to enable eachpath in the network to focus on one forensics operation.

• We perform extensive experiments on challengingdatasets to evaluate our global contrast enhancement de-tection method against existing methods on JPEG images.

II. OUR METHOD

A. Multi-Path Network

The network structure we use is inspired from the VGGmodel [6]. As shown in Fig. 2(a), the VGG network has fivefeature extraction modules including two convolutional layersfollowed by the max-pooling layer. Then, two fully connectedlayers are added to the end of the network.

Different from the original VGG model that uses an imageas input, the input to our network receives the histogram of theimage with the size of 256×1. Since global contrast enhance-ment affects through pixel histograms, the feature extractedfrom the histogram can generally represent the manipulationtype. Moreover, it only requires fewer computation and storage

resources to process histograms than full images. As illustratedin Fig. 2(b), the proposed network consists of three parts:

• Shared layers. The shared layers contain eight convolu-tional layers (conv1 1-conv4 2), and are used to captureshared low-level features of generic image contrast en-hancement operation.

• Operation-specific layers. Operation-specific layers con-sists of several paths, each of which is constructed bytwo convolutional layers and one fully connected layer(conv5 1-conv5 2, fc1). Different contrast enhancementoperation denotes different path without shared weights.

• Aggregation layers. The aggregation layers learn to com-bine the outputs of preceding paths into a discriminativefeature representation using a concatenation layer and afully-connected layer (fc2). Then the binary classificationbetween altered and unaltered images is conducted bya softmax loss. Specifically, we minimize the averageloss E between the true class labels (i.e., unaltered andaltered). The network outputs using the following lossfunction is defined as

E =1

N

N∑i=1

c∑k=1

`ki log(yki ) +

λ

2||ω||2, (1)

where the first term is the cross entropy loss, and thesecond term is the L2 regularization to prevent over-fittingof aggregation layers. `ki and yki are the true label and thenetwork output of the i-th image at the k-th class withN training images and c neurons in the fully-connectedlayer, respectively. Parameter λ is the balancing factorand ω is the vectored weights of the fc2 layer.

Fig. 3. Data augmentation.

B. Data Augmentation

We first briefly describe the way we generate training data.In this work, the positive samples are defined as the manip-ulated images, while the negative samples are the unalteredimages. For data augmentation, we randomly select the contentof each raw image to generate 5 cropped images as the positiveexamples. Let W be the width and H be the height of theraw image. As shown in Fig. 3, we first set (x, y) as thecoordinate of the left top point of the cropped image, in whichx is randomly chosen between 1 and W/4, and y is randomlychosen between 1 and H/4. We randomly choose the widthof the cropped image w between 50 and W − x. To ensurethe aspect ratios of the cropped images, we randomly choosethe height of the cropped region between 0.7w and 1.3w.

We generate negative examples based on positive examples.We first randomly choose the image enhancement operationtype, in this work we consider the gamma correction andthe histogram equalization. The histogram equalization hasno parameters. The parameter of the gamma correction israndomly chosen between 0.4 and 2.1 with the step 0.1, i.e.,[0.4, 0.5, · · · , 2.1]. Then the processed images are stored inthe JPEG format with compression quality 95. All the aboveparameters are set empirically.

C. Multi-stage Training

The back-propagation training algorithm in deep learningdoes not distinguish the operation-specific layers from eachother, which is hard to learn specific representations for thecorresponding forensics operation in each path. Therefore,we develop a three-stage training method to deal with thisproblem. The overall training process is summarized in Algo-rithm 1.Stage 1: training shared layers. The first eight convolutionallayers (i.e., conv1 1-conv4 2 in Fig. 2(b)) are used to captureshared information of histogram under different forensicsoperations. To determine the weights, we add the rest part ofVGG model in Fig. 2(a) to construct a one-path network andtrain it using back-propagation strategy. After several epoches,the shared convolutional layers are determined.Stage 2: training operation-specific layers. Once the sharedlayers are trained, we propose to train two paths individually.

Algorithm 1 Multi-stage training method.Input: altered and unaltered samplesOutput: multi-path network

1: Given all kinds of altered and unaltered samples, we train theVGG model in Fig. 2(a) and fix the bottom layers (conv1 1-conv4 2);

2: for each contrast enhancement operation do3: We remove the other paths of operation-specific layers and

initialize the weights of the current path (conv5 1-conv5 2,fc1) using MPNet in Fig. 2(b);

4: Given each kind of altered and unaltered samples, we train theoperation-specific layers using MPNet in Fig. 2(b);

5: end for6: We enable all the operation-specific layers in MPNet in Fig. 2(b);7: Fixing the previous layers (conv1 1-fc1), we train the aggregation

layers (fc2) using MFNet in Fig. 2(b).

First, with fixed shared layers, we train each operation-specificlayer by removing the other counterparts. Thus the multi-path network degenerates into one-path network, namely VGGnetwork in Fig. 2(a). Then, the operation-specific layer is ini-tialized with random weights. Finally, the network is updatedbased on a mini-batch that consists of the training samplesfrom positive samples under each forensics and negativesamples. The procedure is not finished until the network isconverged.Stage 3: training aggregation layers. After obtaining theoperation-specific layers, we combine them to learn a dis-criminative feature representation. As shown in Fig. 2(b), weconcatenate the operation-specific paths and learn their ag-gregation weights by fully-connected layer (i.e., fc2). The lastsoftmax layer is used for altered/unaltered image classificationusing the loss in (1).

III. EXPERIMENTS

To evaluate the effectiveness of the proposed method, exten-sive experiments are performed and we compare our methodto state-of-the-art Stamm et al.’s method [2]1. Besides, wetrain two SVM models [7] (including linear and RBF kernel,denoted as SVM Linear and SVM RBF) as baselines.Datasets. The Dresden dataset2 [8] is used to evaluate theabove methods. It consists of nature images, dark/flatfieldframes and JPEG scene captured in various indoor and outdoorscenes. We randomly select 16, 000 nature images for training,401 nature images for validation and 1, 851 JPEG sceneimages for testing. We do not select dark/flatfield frames inthe experiment because they include no semantic objects. Todemonstrate the generalization ability of the method, we alsouse the RAISE dataset [9] and the Uncompressed ColourImage Database (UCID) [10] for testing. The RAISE dataset3

contains 5, 999 uncompressed high-resolution images, whichare guaranteed to be camera-native. The UCID dataset4 in-

1We do not compare with Cao et al.’s method [3] because it is failed todetect gap numbers based on the altered image after JPEG compression.

2http://forensics.inf.tu-dresden.de/ddimgdb/publications/ddimgdb3http://mmlab.science.unitn.it/RAISE/download.html4http://jasoncantarella.com/downloads/ucid.v2.tar.gz

cludes 886 available uncompressed images on various top-ics such as natural scenes, man-made objects, indoors andoutdoors. Using the raw images, we generate contrast en-hanced images as described previously. The images are firsttransformed using contrast enhancement operations and thencompressed with a quality factor QF .Metric. To evaluate the global contrast enhancement detectionmethods, each test image is classified by determining if it iscontrast-enhanced or not using a series of decision thresholds.We evaluate them by measuring the true positive rate andthe false positive rate. The Receiver Operating Characteristic(ROC) curves are also generated to calculate the Area Underthe ROC Curve (AUC) score for ranking them. Moreover, wecalculate the probabilities of detection (Pd) and false alarm(Pfa) determined by thresholds (i.e., 0.01, 0.05, 0.1) as thepercentage of the enhanced images correctly classified and thatof the unaltered images incorrectly classified, respectively.Implementation Details. Similar to [2], the green channelof each testing image for training and testing. We compressimages by applying the Python function imwrite in the cv2toolbox at different quality factors. When training our CNN,we set the batch size equal to 256 and optimize the parametersof the network with the Adadelta strategy [11]. We use Xavieralgorithm [12] for weight initialization. We shuffle the trainingset between epochs, and use the early stopping strategy duringthe training process. The balancing factor in (1) is set asλ = 0.01. The learning rate is initially set as 0.4, and decayswhen the accuracy of the validation set is converged. We trainshared layers in 20 epochs. The network is implemented usingTensorflow5 on a machine with a 3.50 GHz Intel i7 5930Kprocessor and 48 GB memory and a NVIDIA GTX 1080Graphical Card.

A. Performance Comparison

As shown in Fig. 4, we present the performance of dif-ferent contrast enhancement (i.e., histogram equalization andgamma correction) detection of our MPNet method and othercompared methods on the Dresden-test, RAISE and UCIDdatasets. MPNet achieves the best AUC scores in identifyingJPEG compression with QF = 95 compared to existingmethods in three datasets. Moreover, our algorithm achieveshigher detection rate Pd > 0.6 even under low Pfa = 0.01.This is attributed to the ability of our network to exploitrelations among several consecutive bins in histogram. Onthe other hand, the SVM models with histogram input fail toconsider such relations, leading to inferior performance. Be-sides, Stamm et al.’s method [2] detects contrast enhancementwithout training, so that it is hard to extract discriminativefeatures from histograms in noisy situations.

B. Discussion

We further perform experiments to study the influence ofvarious important factors of MPNet on the performance.

5We make the source codes of our method and the experimental resultsavailable on our website: https://sites.google.com/site/daviddo0323/.

TABLE ICOMPARISON BETWEEN HISTOGRAM-BASED AND IMAGE-BASED VGG

MODELS IN UCID DATASET WITH QF = 95.

AUC score Speed # of Param. memoryVGG-Hist 93.50 2.03s 3.11M 0.70 MB

VGG-Image 94.14 2.60s 9.73M 59.27 MB

Effectiveness of histogram input. To confirm the assumptionthat the histogram can represent the significant informationof the image, we separately use the histogram and the im-age samples to train the VGG network. In terms of image-based VGG, we crop the image samples with the size of224× 224 randomly based on the resized image with the sizeof 256 × 256. According to the results in Table I, the twomodels achieve comparable performance, but histogram-basedVGG obtains considerable improvement on time and memoryexpenses per sample.Influence of JPEG quality. We explore the performanceof the proposed method with different JPEG quality. In theexperiment, we use the trained images with QF = 95 andtest the images with QF = 90, 70, 50, 30. For comprehensiveevaluation, the ROC curves on different forensics operationsand QFs are presented in Fig. 5. With the compression degreedeeper, the result still shows this method can be instructional.For the middle/low quality factor (QF ≤ 50), our methodkeeps stable and outperforms other methods in a large mar-gin. Stamm et al.’s method [2] achieves less than 50 AUCscore because of failures in detecting peak-gap features whenQF ≤ 50. These results show MPNet performs well on thedetection of the global contrast enhancement operation evenwhen the image is stored in the JPEG format with middle/lowquality.Effectiveness of multi-path network. To demonstrate theeffectiveness of multi-path network, we implement severalVGG models with different output labels, namely VGG-2,VGG-2D, VGG-3 and VGG-3D. Specifically, VGG-2 andVGG-3 correspond to the network with the output labelbeing altered/unaltered and gamma/histogram/unaltered, re-spectively. VGG-2D and VGG-3D indicate the number ofweights from conv5 1 to fc1 is doubled to compare with thetwo paths in our method. As shown in Table II, the multi-path network improves the accuracy moderately. Moreover,our method provides higher detection rate with the samenumber of false alarms. Besides, the performance comparisonbetween VGG-2 and VGG-3 indicates that it is hard to learnefficient discriminative representation of different forensicsoperation by just multi-label output configuration. Meanwhile,increasing the number of parameters of the network bringsa little performance improvement. To sum up, the proposedmulti-path scheme is able to accurately detect several typesof single contrast enhancement operations using the samenetwork architecture and number of parameters.Effectiveness of aggregation layers. The aggregation layersare used to combine the multi-path representation. We remove

Fig. 4. ROC curves of compared methods on Dresden-test, RAISE and UCID datasets with QF = 95. We present the AUC score in the legend.

Fig. 5. ROC curves of compared methods on the (a) Dresden-test, (b) RAISE and (c) UCID datasets with different quality factors (QF = 90, 70, 50, 30).We present the AUC score in the legend.

the fully-connected layer and output the label based on themean softmax score among all the paths, denoted as MPNet-mean. From the results in Table II, the aggregation layers canlearn the discriminative feature integrated from multiple pathsto detect contrast enhancement effectively.

IV. CONCLUSION

In this paper, we propose a new method to detect globalcontrast enhancement based on a deep multi-path network.The network can exploit more discriminative features in con-secutive bins to represent peaks and gaps, even when theoriginal peak and gap features have been diminished by JPEGcompression. Moreover, the multi-path module can furtherimprove the accuracy of forensics detection in different com-

pression quality. Experimental results on the Dresden, RAISEand UCID datasets show the proposed deep model works wellfor contrast enhancement detection, especially when the imageundergos JPEG compression with middle/low quality. In thefuture work, there are several directions for research. First, weplan to explore the performance of our method against anti-forensic techniques [13], [14]. Second, we would expand ourmethod to detect local contrast enhancement.

ACKNOWLEDGMENT

This work was supported in part by the National NaturalScience Foundation of China under Grant 61472388 and Grant61771341, in part by the US Defense Advanced ResearchProjects Agency under Grant FA8750-16-C-0166.

TABLE IITHE PERFORMANCE COMPARISON ON THE DRESDEN-TEST, RAISE AND UCID DATASETS FOR DIFFEREN QUALITY FACTORS (i.e., QF = 90, 70, 50, 30).

THE BEST PERFORMER IS HIGHLIGHTED IN BOLD FONT.

JPEG QF = 90 JPEG QF = 70

Dresden-test MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D Dresden-test MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D

AUC score 93.35 88.03 93.83 94.24 92.54 93.75 AUC score 88.37 80.96 85.53 85.25 84.74 83.02

Pd(Pfa = 0.01) 0.6283 0.2141 0.5815 0.7176 0.5436 0.6797 Pd(Pfa = 0.01) 0.4358 0.1192 0.2919 0.4777 0.2901 0.3588

Pd(Pfa = 0.05) 0.7210 0.6138 0.7388 0.7712 0.7338 0.7662 Pd(Pfa = 0.05) 0.5977 0.4314 0.5073 0.5547 0.5122 0.4911

Pd(Pfa = 0.1) 0.8326 0.7227 0.8281 0.8326 0.8002 0.8186 Pd(Pfa = 0.1) 0.7333 0.5910 0.6350 0.6317 0.6066 0.5943

RAISE MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D RAISE MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D

AUC score 81.67 76.08 81.17 79.57 79.49 80.12 AUC score 78.13 73.34 75.56 73.43 75.25 74.85

Pd(Pfa = 0.01) 0.4146 0.3328 0.2924 0.0101 0.2178 0.3225 Pd(Pfa = 0.01) 0.2285 0.2812 0.1289 0.0098 0.1104 0.1055

Pd(Pfa = 0.05) 0.5145 0.4615 0.4849 0.1268 0.4682 0.4833 Pd(Pfa = 0.05) 0.4023 0.4062 0.3555 0.0605 0.3398 0.3301

Pd(Pfa = 0.1) 0.5848 0.5179 0.5709 0.3268 0.5625 0.5631 Pd(Pfa = 0.1) 0.5176 0.4395 0.4805 0.1816 0.4707 0.4648

UCID MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D UCID MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D

AUC score 83.63 81.17 81.55 81.27 80.30 81.23 AUC score 84.13 81.63 80.95 77.55 79.92 80.53

Pd(Pfa = 0.01) 0.5417 0.5221 0.4896 0.0238 0.4596 0.4831 Pd(Pfa = 0.01) 0.5039 0.4974 0.4323 0.0077 0.3646 0.3737

Pd(Pfa = 0.05) 0.6003 0.5807 0.5859 0.1190 0.5859 0.5820 Pd(Pfa = 0.05) 0.5885 0.5898 0.5573 0.0913 0.5885 0.5352

Pd(Pfa = 0.1) 0.6419 0.6341 0.6289 0.2751 0.6549 0.6367 Pd(Pfa = 0.1) 0.6536 0.6445 0.6042 0.1992 0.6328 0.5977

JPEG QF = 50 JPEG QF = 30

Dresden-test MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D Dresden-test MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D

AUC score 84.88 77.33 76.17 75.23 75.24 74.63 AUC score 92.35 85.16 88.48 89.03 88.59 87.64

Pd(Pfa = 0.01) 0.2444 0.0444 0.1083 0.3477 0.1074 0.2556 Pd(Pfa = 0.01) 0.4135 0.0856 0.3114 0.4860 0.3357 0.4007

Pd(Pfa = 0.05) 0.4453 0.2684 0.2796 0.4118 0.3052 0.3956 Pd(Pfa = 0.05) 0.6200 0.3689 0.5273 0.5820 0.5709 0.5530

Pd(Pfa = 0.1) 0.6490 0.4939 0.4364 0.4916 0.4375 0.4749 Pd(Pfa = 0.1) 0.8013 0.6283 0.6819 0.7026 0.7081 0.6685

RAISE MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D RAISE MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D

AUC score 73.68 69.28 66.07 64.13 65.94 63.89 AUC score 82.31 74.95 78.23 74.96 78.93 74.96

Pd(Pfa = 0.01) 0.1602 0.1248 0.0858 0.0037 0.0822 0.0948 Pd(Pfa = 0.01) 0.3145 0.2128 0.2843 0.0066 0.2572 0.0066

Pd(Pfa = 0.05) 0.3789 0.2977 0.2441 0.0229 0.2420 0.2824 Pd(Pfa = 0.05) 0.4963 0.3825 0.4531 0.0387 0.4480 0.0387

Pd(Pfa = 0.1) 0.4635 0.3956 0.3412 0.0798 0.3599 0.3645 Pd(Pfa = 0.1) 0.5895 0.4701 0.5372 0.1424 0.5712 0.1424

UCID MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D UCID MPNet MPNet-mean VGG-2 VGG-3 VGG-2D VGG-3D

AUC score 85.80 83.49 82.36 80.38 80.67 81.78 AUC score 92.07 85.17 89.46 86.88 90.02 88.67

Pd(Pfa = 0.01) 0.5273 0.5365 0.3711 0.0078 0.3841 0.3060 Pd(Pfa = 0.01) 0.5977 0.5547 0.5013 0.0119 0.5391 0.4297

Pd(Pfa = 0.05) 0.6302 0.6328 0.5781 0.0727 0.5638 0.5273 Pd(Pfa = 0.05) 0.7396 0.6862 0.6706 0.1675 0.7057 0.6484

Pd(Pfa = 0.1) 0.6849 0.6927 0.6484 0.2701 0.6302 0.6237 Pd(Pfa = 0.1) 0.8060 0.7578 0.7487 0.5198 0.7956 0.7370

REFERENCES

[1] G. Cao, Y. Zhao, and R. Ni, “Forensic estimation of gamma correctionin digital images,” in ICIP, 2010, pp. 2097–2100.

[2] M. C. Stamm and K. J. R. Liu, “Forensic detection of image manip-ulation using statistical intrinsic fingerprints,” TIFS, vol. 5, no. 3, pp.492–506, 2010.

[3] G. Cao, Y. Zhao, R. Ni, and X. Li, “Contrast enhancement-basedforensics in digital images,” TIFS, vol. 9, no. 3, pp. 515–525, 2014.

[4] L. Wen, H. Qi, and S. Lyu, “Contrast enhancement estimation for digitalimage forensics,” TOMM, 2018.

[5] T. Pevny and J. J. Fridrich, “Detection of double-compression in JPEGimages for applications in steganography,” TIFS, vol. 3, no. 2, pp. 247–258, 2008.

[6] K. Simonyan and A. Zisserman, “Very deep convolutional networksfor large-scale image recognition,” CoRR, vol. abs/1409.1556, 2014.[Online]. Available: http://arxiv.org/abs/1409.1556

[7] S. Shalev-Shwartz, Y. Singer, N. Srebro, and A. Cotter, “Pegasos: primalestimated sub-gradient solver for SVM,” Math. Program., vol. 127, no. 1,pp. 3–30, 2011.

[8] T. Gloe and R. Bohme, “The ’dresden image database’ for benchmarkingdigital image forensics,” in ACM SAC, 2010, pp. 1584–1590.

[9] D. Dang-Nguyen, C. Pasquini, V. Conotter, and G. Boato, “RAISE: araw images dataset for digital image forensics,” in ACM MMSys, 2015,pp. 219–224.

[10] G. Schaefer and M. Stich, “UCID: an uncompressed color imagedatabase,” in Storage and Retrieval Methods and Applications forMultimedia, 2004, pp. 472–480.

[11] M. D. Zeiler, “ADADELTA: an adaptive learning rate method,” CoRR,vol. abs/1212.5701, 2012. [Online]. Available: http://arxiv.org/abs/1212.5701

[12] X. Glorot and Y. Bengio, “Understanding the difficulty of training deepfeedforward neural networks,” in AISTATS, 2010, pp. 249–256.

[13] G. Cao, Y. Zhao, R. Ni, and H. Tian, “Anti-forensics of contrastenhancement in digital images,” in MM&Sec, 2010, pp. 25–34.

[14] C. W. Kwok, O. C. Au, and S. H. Chui, “Alternative anti-forensicsmethod for contrast enhancement,” in IWDW, 2011, pp. 398–410.